A recent study challenges the prevailing view that large language models are merely statistical parrots, instead proposing that Transformers actively construct and apply prompt-dependent transformations during inference. This computation, termed Sequence-level Interactive Dynamic Processing (SIDPP), enables the model to generate parameters on the fly, allowing for more dynamic and interactive processing. The researchers argue that this interpretation is more accurate, as it accounts for the model's ability to generalize and respond to novel inputs. The SIDPP framework has significant implications for the development of large language models, as it suggests that these models are capable of more complex and nuanced processing than previously thought1. This, in turn, has important security implications, as the increased capability of these models also expands their potential risk surface, making it essential for practitioners to reassess their security protocols.